There is no good way to tell it "this isn't a conversation, just repeat the answer to the initial prompt again".
The solution is to just re-paste the initial prompt in each time, but still it isn't ideal. There isn't a good way to tell chatgpt "you can throw away all the context after the initial prompt and up until now".
Of course the entire point of ChatGPT is that it maintains a conversation thread, so I get why they don't fix up this edge case.
My problem is more of, I give ChatGPT some complicated instructions, and it'll start forgetting the early on instructions long before any token limit is reached.
So for example, if early on I ask for certain tokens to be returned in parens, well my initial prompt is too long, it'll forget the parens thing and start returning tokens without the surrounding (), which then breaks my parser!
As she kept asking for more, I prompted "great, do another one" and eventually my original instruction fell out of the context window. It continued to generate a children's story, but with no more blanks.